基于稀疏表示的图像重建与去噪方法研究

发布时间：2018-05-28 01:30

本文选题：压缩感知 + 稀疏表示　；参考：《湖北工业大学》2017年硕士论文

【摘要】：随着图像信息需求量的与日俱增,在信息采样与传输过程中,传统奈奎斯特(Nyquist)定理所带来的抽样资源浪费、硬件成本昂贵、信息处理效率低下等局限性问题日益突出,而以信号稀疏性为前提的压缩感知(Compressed Sensing,CS)采样编码技术,就避免了大量冗余数据的产生,有效地提高了信号处理的效率,并且降低了对设备采样速率的要求,极大地节省了数据存储空间和传输成本。本论文结合CS理论,深入研究基于稀疏表示的图像重构以及去噪方法,将稀疏表示模型用于图像的传输、编码以及污染图像的去噪应用当中,以求高效率、高精准度地恢复出完整图像信号。论文主要在如下几个方面取得了研究成果。(1)针对图像混合噪声去除不足问题,提出一种图像块分组的加权编码方法来改善图像去噪质量。首先,从训练图像中利用非局部相似块来提取出块分组;然后,用得到的块分组来训练非局部自相似先验模型;最后,集成稀疏先验模型和非局部自相似先验模型到正则化项和编码框架中。实验表明,用此方法所得的去噪图像峰值信噪比较同类方法提高了0.036~2.865dB,获得了更好的图像去噪效果。(2)针对腐化图像恢复不足问题,给出一种基于PCA的非局部聚类稀疏表示模型来提高恢复质量。首先,用图像非局部自相似性来取得稀疏系数值;然后,对观测图像的稀疏编码系数进行集中聚类;最后,通过学习字典使降噪图像的稀疏编码系数接近原始图像的编码系数。实验表明,所提方法所得的重建图像峰值信噪比较同类方法平均提高了0.5653 dB,获得了更好的图像重建质量。(3)针对图像修复技术缺陷,设计出一种高斯尺度训练稀疏表示方法以达到高分辨率重建效果。首先,利用非局部相似块提取出分组的块群;然后,利用同步稀疏编码得到非局部扩展高斯尺度混合模型;最后,将块分组模型和高斯尺度稀疏模型联合到编码框架中。实验表明,该方法既能保留图像的边缘又能抑制人工操作造成的不利影响,重建出的图像峰值信噪比较其他同类竞争方法提高了0.02~0.64 dB。
[Abstract]:With the increasing demand of image information, in the process of information sampling and transmission, the traditional Nyquist (Nyquist) theorem is a waste of sampling resources, high cost of hardware and low efficiency of information processing, and the Compressed Sensing (CS) sampling and coding technique based on signal sparsity It avoids the production of a large number of redundant data, effectively improves the efficiency of signal processing, and reduces the requirement for the sampling rate of the equipment, greatly saves the data storage space and transmission cost. This paper studies the image reconstruction and denoising based on the sparse representation based on the CS theory, and uses the sparse representation model to use the sparse representation model. In the application of image transmission, encoding and de-noising of contaminated images in order to efficiently and accurately restore the complete image signal. The main research results are obtained in the following aspects. (1) a weighted coding method for image block grouping is proposed to improve the quality of image denoising in view of the lack of image mixed noise removal. First, the block grouping is extracted from the non local similar blocks in the training image. Then, the non local self similar prior model is trained by the block grouping. Finally, the sparse prior model and the non local self similar prior model are integrated into the regularization term and the coding frame. The experimental results show that the peak signal to noise ratio of the denoised image obtained by this method is shown. Compared with the same method, 0.036~2.865dB improves the image denoising effect. (2) to improve the recovery quality, a non local clustering sparse representation model based on PCA is given to improve the recovery quality. First, the sparse coefficients are obtained by the non local self similarity of the image, and then the sparse coding coefficients of the observed images are obtained. In the end, the sparse coding coefficient of the denoised image is close to the coding coefficient of the original image by learning dictionary. The experiment shows that the peak signal to noise ratio of the reconstructed image is improved by 0.5653 dB, and the quality of the image reconstruction is better. (3) for the defect of image repair technology, a design is designed. The Gauss scale training sparse representation method is used to achieve high resolution reconstruction effect. First, the block groups are extracted from the non local similar blocks; then, the non local extended Gauss scale hybrid model is obtained by the synchronous sparse coding. Finally, the block grouping model and the Gauss ulnar sparse model are combined into the coding framework. The experiment shows that this method is applied to the coding framework. The method not only preserves the edge of the image, but also inhibits the adverse effect caused by manual operation. The reconstructed image peak signal to noise comparison of other similar competition methods improves 0.02~0.64 dB.
【学位授予单位】：湖北工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】